Mukayese: Turkish NLP Strikes Back

Summarization: mukayese/transformer-turkish-summarization

This model is uncased, it was initialized from scratch and trained only the mlsum/tu dataset with no pre-training.

It achieves the following results on the evaluation set:

• Rouge1: 43.2049
• Rouge2: 30.7082
• Rougel: 38.1981
• Rougelsum: 39.9453

Check this paper for more details on the model and the dataset.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 4
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 2
• total_train_batch_size: 64
• total_eval_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 15.0
• mixed_precision_training: Native AMP
• label_smoothing_factor: 0.1

Framework versions

• Transformers 4.11.3
• Pytorch 1.8.2+cu111
• Datasets 1.14.0
• Tokenizers 0.10.3

Citation

@misc{safaya-etal-2022-mukayese,
    title={Mukayese: Turkish NLP Strikes Back},
    author={Ali Safaya and Emirhan Kurtuluş and Arda Göktoğan and Deniz Yuret},
    year={2022},
    eprint={2203.01215},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}